Index: sys/kern/vfs_subr.c
===================================================================
--- sys/kern/vfs_subr.c
+++ sys/kern/vfs_subr.c
@@ -6022,8 +6022,12 @@
 		 * Try-lock because this is the wrong lock order.  If that does
 		 * not succeed, drop the mount vnode list lock and try to
 		 * reacquire it and the vnode interlock in the right order.
+		 *
+		 * We VI_TRYLOCK_CONTENDED since with the sheer number of
+		 * vnodes to be inspected quite a few of them can already have
+		 * the interlock held if the machine is under load.
 		 */
-		if (!VI_TRYLOCK(vp) &&
+		if (!VI_TRYLOCK_CONTENDED(vp) &&
 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
 			goto restart;
 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
Index: sys/sys/vnode.h
===================================================================
--- sys/sys/vnode.h
+++ sys/sys/vnode.h
@@ -438,6 +438,19 @@
 #define	VI_LOCK(vp)	mtx_lock(&(vp)->v_interlock)
 #define	VI_LOCK_FLAGS(vp, flags) mtx_lock_flags(&(vp)->v_interlock, (flags))
 #define	VI_TRYLOCK(vp)	mtx_trylock(&(vp)->v_interlock)
+/*
+ * To be used when the OP is useful to do, but likely enough to fail that it
+ * makes sense to pre-read the lock first (and consequently avoid dirtying
+ * the cache line).
+ */
+#define	VI_TRYLOCK_CONTENDED(vp) ({				\
+	int _rv;						\
+	if (mtx_owner(VI_MTX(vp)) != NULL)			\
+		_rv = 0;					\
+	else							\
+		_rv = mtx_trylock(&(vp)->v_interlock);		\
+	_rv;							\
+})
 #define	VI_UNLOCK(vp)	mtx_unlock(&(vp)->v_interlock)
 #define	VI_MTX(vp)	(&(vp)->v_interlock)